类型相似性与品味影响力的推荐系统评分预测

Rating Prediction in Recommender Systems Based on Type Similarity and Taste Influence

电影推荐算法依靠计算用户的偏好差异进行推荐,以解决互联网时代的信息过载问题.传统协同过滤等推荐算法主要基于用户间对相同电影的评分差异计算用户偏好的相似性.这类方法忽视了用户的评分行为是一种实际上的选择行为,即便评分不高也体现出用户对该类型电影的兴趣.针对这一问题,本文设计了基于电影类型标签选择概率的用户间相似性计算方法,并建立了以用户为节点,以用户之间的相似性为边的推荐系统的复杂网络模型,并根据上述网络拓扑结构中的节点中心性数据,进一步设计了平衡用户品味影响力函数,调整了用户协同偏好的结果,提出了基于用户偏好相似性和用户品味影响力的电影评分预测方法.在MovieLens数据集上的实验结果表明,本文提出的算法与几种典型的现有方法相比较,可以有效的度量用户偏好的相似性以及抵消用户大众化品味影响力被高估在评分预测中带来的负面影响,与现有算法相比预测误差平均降低了2%至5%.

Movie recommendation algorithm relies on calculating the user’s preference difference to recommend,in order to solve the problem of information overload in the Internet era. Traditional recommendation algorithms such as collaborative filtering mainly calculate the similarity of user preferences based on the difference of ratings of the same movie between users. This kind of method ignores the user’s rating behavior,which is a kind of actual choice behavior,even if the rating is not high,it also reflects the user’s interest in this type of film. To solve this problem,this paper designs a similarity calculation method between users based on the selection probability of movie type tags,and establishes a complex network model of the recommender system with users as the nodes and the similarity between users as the edges. According to the node centrality in the network topology structure,a function to balance user’s taste influence is designed to adjust user collaboration based on the similarity of user preferences and the influence of user taste,a method of movie rating prediction is proposed. The experimental results on MovieLens dataset show that the proposed algorithm can effectively measure the similarity of user preferences and offset the negative impact of overestimation of users’ taste influence in rating prediction compared with several typical existing methods. The prediction error of this algorithm is reduced by 2% to 5% on average compared with the existing algorithms.